llama.cpp lands Qwen3-Next / Qwen3.5 conversion metadata, a QKV fusion flag, and two CUDA race fixes in 36 hours

GitHub·medium signalXBlueskyLinkedInCopy link

ggml-org/llama.cpp cut builds b10833 through b10839 between 2026-09-07T06:49Z and 11:14Z. #28208 writes explicit `recurrent_layers` during Qwen3-Next / Qwen3.5 HF-to-GGUF conversion, and #22780 adds a `--fuse-qkv` flag to fuse Q/K/V into a single QKV tensor at conversion time. Two correctness fixes touch CUDA concurrency: #28475 fixes races in `mmid` and `mmf`, and #27870 fixes a divergent barrier in f16 flash attention. Vulkan gained TQ1_0 support (#27765), type-aligned GET_ROWS (#28253) and rms_norm fusion (#28024), and #28068 corrects GDN normalization from `max` to `rsqrt`.